Machine Learning in Computer Architecture and Systems (MLArchSys)

  • LECTURER: Thaleia Dimitra Doudali
  • AFFILIATION: IMDEA Software Institute

Outline

Modern applications generate huge amounts of data and exhibit more irregular data access patterns that break the effectiveness of traditional hardware and software systems. The integration of Machine Learning (ML) methods has the potential to accelerate such systems via intelligent, robust and adaptive decision making. However, the necessity, practicality, effectiveness and interpretability of machine learning-based systems is still ambiguous. This pilot seminar series will cover an overview and three concrete use cases of machine learning deployed to improve hardware- and operating system-level data management of cache and memory computer resources.

Syllabus

  1. Introduction. Overview of Systems Challenges for Artificial Intelligence.
  2. ML for cache replacement. ML method: Reinforcement Learning
  3. ML for data prefetching. ML method: Long Short-Term Memory Networks (LSTMs) - Classification.
  4. ML for hybrid memory management. ML method: Long Short-Term Memory Networks (LSTMs) - Regression.

Assessment Method

Short report on the takeaways of each class, answering predefined questions. The report should be emailed to the Professor until the next class (1 week). Each report accounts for 25% of the overall grade, since there will be 4 reports in total.

Prerequisites

Basic understanding of computer systems and architecture, although each class will include an overview of the problem, for a general audience. Similarly, each class will summarize the functionality of the discussed machine learning method, no ML expertise is required.

Lective hours

6

Remarks

The seminar will be comprised of 4 lectures, 2 hours each, over the span of 4 weeks.

Structure of each class:

  • Overview of the problem.
  • Overview of existing non-ML solutions.
  • Overview of the ML-based method and proposed solutions.
  • Class discussion on ML necessity, practicality, effectiveness and interpretability

Recommended Reading

  • Overview: A Berkeley View of Systems Challenges for AI
  • ML for cache replacement: Designing a Cost-Effective Cache Replacement Policy using Machine Learning
  • ML for data prefetching: Learning Memory Access Patterns
  • ML for hybrid memory management: Kleio: a Hybrid Memory Page Scheduler with Machine Intelligence

Timetable

  • 14 March, 19:00-21:00
  • 21 March, 19:00-21:00
  • 28 March, 19:00-21:00

Lecture Theatre

  • A-6306

Tuition Language

English.